A Memory-Free Evolving Bipolar Neural Network for Efficient Multi-Label Stream Learning

Abstract

Many fields, like document tagging, video labeling, and medical analysis, require associating the samples with multiple non-exclusive labels, driving the research in multi-label learning. Unlike several multi-label learning setups, practical applications are challenging because they need learning from a stream of samples and labels. This work proposes an Evolving Bipolar Network architecture called EBN-MSL consisting of two parallel layers trained in a maximum margin framework to learn efficiently in a continual multi-label learning scenario without utilizing any samples stored from previous tasks. This work considers two learning setups, one for separately learning each task (SEA*) and another for jointly learning subsequent tasks (SEA). Experiments on benchmark multi-label learning datasets establish the superior learning capability of EBN-MSL in the presence of samples having all positive or negative labels. Results indicate that EBN-MSL (both SEA and SEA* setups) significantly outperforms the current state-of-the-art architecture-based continual multi-label learning algorithm.